International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.

4, October 2014

THE N-DIMENSIONAL MAP MAKER ALGORITHM

John R Rankin
Department of Computer Science and Computer Engineering
La Trobe University, Australia

ABSTRACT
The Map Maker algorithm which converts survey data into geometric data with 2-dimensional Cartesian
coordinates has been previously published. Analysis of the performance of this algorithm is continuing. The
algorithm is suitable for generating 2D maps and it would be helpful to have this algorithm generalized to
generate 3D and higher dimensional coordinates. The trigonometric approach of the Map Maker algorithm
does not extend well into higher dimensions however this paper reports on an algebraic approach which
solves the problem. A similar algorithm called the Coordinatizator algorithm has been published which
converts survey data defining a higher dimensional space of measured sites into the lowest
dimensionalcoordinatization accurately fitting the data. Therefore the Coordinatizator algorithm is not a
projection transformation whereas the n-dimensional Map Maker algorithm is.

KEYWORDS
N-dimensional space, Projections, Distance Matrices, Map Maker, Coordinatizations

1. INTRODUCTION
The Map Maker algorithm [1] serves a fundamental purpose. Given geodetic survey data on N
sites Pi for i = 1 to N in the form of an NxN distance matrix D the algorithm generates 2D
coordinates for each point i.e. it determines xi and yi such that Pi = (xi,yi) satisfies the distance
matrix D, according to

 =

(  ,  )

(1)

for all i and j = 1 to N. The points Pi however may not in reality all lie on a single flat plane and
geodetic survey data is well-known to be non-planar. For example we can test the Map Maker
algorithm by starting with a given three dimensional point set Pi = (xi,yi,zi) for i = 1 to N and from
these generate a distance matrix D via Eq(1) using the 3-dimensional Euclidean distance measure.
Then subsequently we may take this distance matrix D and feed it through the Map Maker
algorithm to produce N points Qi for i = 1 to N in 2 dimensions. The generated points Qi can then
be plotted on the flat screen or output page. Taken together these steps constitute a projection of
points Pi in 3D to corresponding points Qi in 2D. This projection is not a formula applicable to
single points as it requires a set of N 3 points to perform the process. The first point P1 is fixed
at the origin so that Q1 = (0,0). The second point P2is placed along the x-axis in the positive
direction so that Q2 = (D12,0). The third point P3 is placed such that Q3 has 2-dimensional distance
D13 from point Q1 and 2-dimensional distance D23 from point Q2 and a positive y component. This
is equivalent to selecting a plane in 3D space which is the plane of the 3D triangle P1P2P3, then
selecting an origin in the plane as point P1, the x-axis direction as the vector P1P2 and the y axis
perpendicular to this such that P3 is on the positive y side of the plane. All other points Pifor i = 4
to Nare projected onto this plane by rotating each triangle PiP1P2in 3D to the triangle PiP1P2
where Pi lies in the plane of P1P2P3. Whether Pi is placed in the positive y half of the plane or
DOI : 10.5121/ijcga.2014.4402

19

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

the negative side is based on which choice has a closer distance value to the distance D3i. In this
way all distances D1i and D2i are preserved and D3i is as close to preserved as possible in the
projection. Denoting the distance matrix of the projected points as D we therefore have D1i = D1i
and D2i = D2i for i = 1 to N. The other elements of the D and D matrices will in general differ
and the Map Maker algorithm tries to minimize these differences. A measure of the effectiveness
of this minimization is the root mean square error (RMSE) in the differences of D and D. Work
in reference [1] suggests that reordering the points which results in a permutation of the D matrix
can result in lower RMSE values. The RMSE values are also dependent on N, the range of
coordinate values that were used to generate D and their dimensionality.
Two other projection operations were proposed by Lee in [2] called the Second Nearest
Neighbour approach and the Reference Point approach although he does not attempt to compute
coordinates for the mapped points. While the Map Maker algorithm preserves the distances of all
points to P1 and P2, the Second Nearest Neighbour approach preserves the distances of the next
mapped point Pk to the two nearest previously mapped points Pi and Pj and the Reference Point
approach preserves distances of the next mapped point Pk to P1 and the previously mapped nearest
point Pi (to point Pk). Thus Lees methods lead to different spatial locations and therefore
coordinatizations of the sites Pi unless the sites are intrinsically two-dimensional from the start.
Other authors ([3-12]) have shown the importance of locating objects in space given the distance
matrix alone. Rankin in [13] has shown the remarkable result that the distance matrix can
determine the dimension of the embedding space and that this space can be determined to be
either Euclidean or pseudo-Euclidean from the distance matrix alone. This is achieved through
the Coordinatizator algorithm described in that paper.
Youldash and Rankin {14] have shown the methodology adopted for evaluating the speed and
accuracy of the Map Maker algorithmand preliminary results were given. Graphing the results
was found to be difficult due to the wide statistical variations coming from random point sets so
that statistical methods need to be applied. While the 2D Map Maker algorithm generates a 2D
coordinatization of the survey sites from their distance matrix, the 3D Map Maker algorithm
generates a 3D coordinatization of the survey sites from their distance matrix. It is therefore
reasonable to expect that if the survey sites are intrinsically located in a 3-dimensional space, that
the 3D Map Maker algorithm should in general produce lower RMSE values. The first aim of this
research was to create the 3D Map Maker algorithm and thence test this conjecture. This is
achieved in section 2 below with the pseudo-code in section 3 and testing results in section 6. The
next aim of the research was to carry this process forward into n-dimensional space and sections 4
and 5 below report on this work.

2. THE 3D MAP MAKER ALGORITHM

Consider N sites Pi with a distance matrix D and we want to create 3D coordinates for each site.
We will proceed as in the Map Maker algorithm by declaring the first site P1 to be located at the
origin so that
 = (0, 0, 0)
(2)

Next P2 is declared to define the direction of the x-axis preserving the distance D12 so that
 = (
 , 0, 0)

(3)

Next site P3 is declared to define the direction of the y axis by having P3y> 0 and P3z = 0. As in the
Map Maker algorithm we compute the angle P2P1P3 as so that
 = (

,
  , 0)
(4)
20

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

Finally we define the sense of the z-axis perpendicular to the plane of P1P2P3 by asserting that P4
has a positive z component. Figure 1 shows the geometry where C is the point subtended by P4 on
the plane of triangle P1P2P3. Thus
 = ( ,  ,  )
(5)

where z4> 0.We need to relate the x4, y4 and z4 to the components of D. Thus the problem is to
solve the tetrahedron for x4, y4 and z4 when all the side lengths D12, D13, D14, D23, D24 and D34 of
the tetrahedron are known. The solution is:

 =








+

2





2



+



 =
3
  

 
 = 


(6)

(7)

(8)

where is the same angle used above andalso used in the 2D Map Maker algorithm.
We now enter a loop to place points Pi for i = 5 to N in 3D space. To place point Pii.e. to
determine xi, yi and zi, we use the same formulas as Equations (6) to (8) except replacing index 4
with i. For placing P4 the positive root is taken in equation (8) but for index i> 4 we must decide
for each point Piwhether to flip the point and use the negative root instead. As with the 2D Map
Maker algorithm, the decision is made by seeing which point Pi or Pi flipped has a distance to P4
which is closer in value to distance D4i.

Figure 1. Deriving the coordinates of Pi for i = 4 to n.

21

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

3. 3D MAP MAKER PSEUDO-CODE

The 3D Map Maker algorithm can be expressed in the following pseudo-code which is readily
converted to C++, Java or other suitable programming language. It assumes only that the NxN
real matrix D is available.
Set P[1] = (0,0,0)
Set P[2] = (D[1,2],0,0)
Set P[3] = (D[1.3]*cos_alpha,D[1,3]*sin_alpha,0)
for i = 4 to N
Set x = (Sq(D[1,2]) Sq(D[2,i]) + Sq(D[1,i]))/(2*D[1.2])
Set y = (Sq(P[3].x) + Sq(P[3].y) 2*P[3].x*x Sq(D[3,i]) + Sq(D[1,i]))/2*P[3].y)
Set z = SqRoot(Sq(D[1,i]) Sq(x) Sq(y))
Set P[i] = (x,y,z)
// test for point flipping:
if i> 4 then
Set Flipped = (x,y,-z)
if Abs(D[4,i] Distance(P[4],Flipped)) < Abs(D[4,i] Distance(P[4],P[i])) then
Set P[i] = Flipped
This pseudo-code assumes that 3D points are stored as C++ structs with components called x, y
and z and that the point array is dimensioned to hold elements 1 to N. The terms cos_alpha and
sin_alpha are computed trigonometrically (once only for all sites i> 4) as in the Map Maker
algorithm. The code has been implemented and undergone extensive testing which is reported in
section 6.

4. THE N-DIMENSIONAL MAP MAKER ALGORITHM

In 2D we used the first 3 points to define the 2D coordinate system: P1 defines the origin of
coordinates, P2 defines the x-axis and P3 defines the y-axis.In 3D we used the first 4 points to
define the 3D coordinate system: P1 defines the origin of coordinates, P2 defines the x-axis, P3
defines the y-axis and P4 defines the z-axis. So in n dimensions we need to use the first n+1 points
to define the coordinate system and then the remaining points can be located inside this
coordinate system. Therefore the first n+1 points will be coordinatized as follows:
 = (0, 0, 0, 0)
 = ( , 0 ,0, 0)
 = ( ,  , 0, 0)
 = (,  ,  , 0)

(
" = " , " , " , " "# , 0)
"$ = ("$  , "$  , "$ , "$ " )

In these formulas it is required in defining axial directions that

 # > 0

(9)

(10)

fori = 2 to n+1. There is no sign restriction on the other xij values. It is also clear that in selecting
a projection dimension n we are restricted to the requirement:
22

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

& +1

(11)

In the case of equality in (11), the problem reduces to the Coordinatizator algorithm solved in
[13]. For the n-dimensional Map Maker algorithm therefore we will assume that n < N-1.The
components xij for i = 2 to n+1 and j = 1 to i-1 are n(n+1)/2 unknowns to be determined first.
After the xij are determined then the components of any point Pi for i> n+1 can be determined in
terms of the xij and D. To see how the components of Pi are determined in terms of the xij and D,
for simplicity denote the components of Pi as xi for i = 1 to n. Then form the n equations:
()* )+ )

.,*+

+#2

"

= /(01 0+1 ) + / 4

,

132

43

(12)

for j = 1 to n. These are n equations in n unknowns, the xi for i = 1 to n. The general solution is
given by the following recurrence relation:
4 =


5/ 4$ 
3



4$ 



 

for k = 1 to n-1 and finally

" =


9
 

4#

2 /  4$  6 /(24$ 4 )
3

(13)

"#

/ 
3

(14)

The plus or minus sign in equation (14) indicates that a flip decision has to be made for this
component to minimize the RMSE. The decision is based on which point flipped or unflipped
results in a distance value closer to Di n+1.Equations (13) and (14) are the general solution to
finding the coordinates of Pi for i> n+1 in terms of the components xij and the given distance
matrix D. Equation (13) is the generalization of the cosine rule as used in Map Maker and
equation (14) is the generalization of the sine formula used in the Map Maker algorithm.
Furthermore this procedure for deriving the coordinates of Pialso applies to finding the xij in the
first place by sequentially solving for the xij down the list given in (9) and always taking the
positive root in equation (14). However in doing this care must be taken in computing only the
coordinates up to i-2 allowing the remainder to be zero. This process is shown in the algorithm
presented in the next section.

5. ALGORITHM
For implementing the n-dimensional Map Makers algorithm the following pseudo-code
represents this new algorithm. For this algorithm each point is a dynamic array of floating point
numbers (doubles) of length n and the set of all points is a dynamic array of length N of this point
type.
for i = 1 to N
for j = 1 to n
Set P[i,j] = 0
23

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

Set P[2,1] = D[1,2]

Set P[3,1] = D[1.3]*cos_alpha
Set P[3.2] = D[1,3]*sin_alpha
for i = 4 to n+1
for j = 1 to n
Pi[j] = 0
for k = 1 to i-2
Set sum1 = 0
Set sum2 = 0
for j = 1 to k
Set sum1 = sum1 + Sq(P[k+1,j])
Set sum2 = sum2 + Pi[j]*P[k+1,j]
Set P[I,k] = (sum1 Sq(D[k+1,i]) + Sq(D[1,i]) 2*sum2)/(2*P[k+1,k])
Set sum1 = 0
for j = 1 to n-1
Set sum1 = sum1 Sq(Pi[j])
Set Pi[i-1] = SqRoot(sum1)
for j = 1 to n
Set P[i,j] = Pi[j]
for i = n+2 to N
for j = 1 to n
Pi[j] = 0
for k = 1 to n-2
Set sum1 = 0
Set sum2 = 0
for j = 1 to k
Set sum1 = sum1 + Sq(P[k+1,j])
Set sum2 = sum2 + Pi[j]*P[k+1,j]
Set P[i,k] = (sum1 Sq(D[k+1,i]) + Sq(D[1,i]) 2*sum2)/(2*P[k+1,k])
Set sum1 = 0
for j = 1 to n-1
Set sum1 = sum1 Sq(Pi[j])
Set Pi[n] = SqRoot(sum1)
for j = 1 to n-1
Set Flipped[j] = Pi[j]
Set Flipped[n] = -Pi[n]
if Abs(D[n+1,i] Distance(P[n+1],Flipped)) < Abs(D[n+1,i] Distance(P[n+1],P[i])) then
Set Pi = Flipped
for j = 1 to n
Set P[I,j] = Pi[j]
The algorithm for the n-dimensional Map Maker is clearly longer and more detailed than either
the Map Maker or 3D Map Maker algorithms and requires more memory. The next section will
present the initial results and implications of this algorithm.

24

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

6. DISCUSSION
Normally when making maps from survey distance information we expect that the maps will be
drawn on a flat 2D sheet of paper or flat screen and the Map Maker algorithm performs this
service. However the concept is generalizable so that given the distance matrix for a set of N sites
one can now generate Cartesian coordinates for each site for any specified dimension n decided in
advance. For geodetic survey data in our 3-dimensional world we could expect that n = 3 is
sufficient. The 3D output when displayed graphically shows the positional relationships between
the survey sites and the user can view this 3D data from many different angles and viewpoints.
Research is underway to convert the 3D coordinatized point data into a height map that can show
the detailed topography and lie of the land.
To test the 3D Map Maker algorithm, many sets of random 3D points were constructed. These
sets were each converted into distance matrices and stored in files. The distance matrix files were
then given to the 3D Map Maker algorithm which converted the distance matrix input to 3D
coordinates for each survey point. The output 3D point sets do not directly compare with the
initial 3D point sets that were constructed because the algorithm causes a translation operation to
P1 as origin and then reflections and rotations of the points about P1so that P1P2 aligns with the xaxis, P3 is in the x-y plane with positive y component and P4 has positive z component. Therefore
instead of comparing the point sets, we compute a new distance matrix based on the output points
from the algorithm and compare this with the original distance matrix. The results varied based
on the number N of points in the set. As an example with N = 5, a total RMS absolute error value
of 0.000104 was obtained with a total RMS relative error of 0.00000581. For N = 10 a sample
result was 0.00145 for the absolute error and 0.00337 for the relative error.
A similar testing process was used on the n-dimensional Map Maker algorithm. For N = 10 and n
= 5 a total RMS absolute error was 0.0056 with a total RMS relative error of 0.00577. Low errors
suggests that the n-dimensional Map Maker algorithm is doing its job correctly. If the distance
matrix used in the algorithm was created from a point set of dimensionality different from n then
we could expect the error rate to rise. This is subject to future testing.
It should be noted that the n-dimensional Map Maker algorithm produces enantiomorphs because
handedness (i.e. parity) cannot be transmitted in the distance matrix. This means that the spatial
arrangement of the nodes can be any of 2n-1 possible mirror images produced by parity inversions
i.e. reflections along any axis other than the first and all enantiomorphs have the same distance
matrix D making them indistinguishable in the Map Maker algorithms.

CONCLUSIONS
The n-dimensional Map Makers algorithm converts spatial distance survey data to point
coordinates in n dimensions. The first n+1 points get mapped to prescribed positions to define the
positive directions of the n axes. In so doing these first n+1 points contain n(n+1)/2 zeros in their
coefficients and preserve n(n-1)/2 distances. If the distance matrix was originally derived not by
survey measurement but from N points in a k-dimensional space then the n-dimensional Map
Makers algorithm can be viewed as a projection transformation from k dimensions to n
dimensions. If N > n+1 and k > n we can expect to see distance preservation failures indicated by
a significant positive RMSE. If N n+1 or k n there should be no significant systemic errors so
that only numerical calculation errors should appear. Continuing work in this area includes testing
of the speed of the algorithm and the behaviour of the RMS errors as well as applications of the
algorithm to real world problems.
25

International Journal of Computer Graphics & Animation (IJCGA) Vol.4, No.4, October 2014

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]

Rankin J, Map Making From Tables IJCGA, April 2013, Vol3, Nr 2 pp 11-17.
http://airccse.org/journal/ijcga/current2013.html
Lee RTC et al, A Triangulation Method for the Sequential Mapping of Points from N-Space to TwoSpace, IEEE Transactions on Computers, March 1977, pp288-292.
Yin H, Nonlinear Dimensionality Reduction and Data Visualization: A Review, International
Journal of Automation and Computing 04(3), July 2007, pp294-303.
Sammon J W Jr, A Nonlinear Mapping for Data Structure Analysis, IEE Transaction on
Computers, Vol C-18, No. 5, May 1969, pp401-409.
Shisko JF et al, IGODS: An Important New Tool for Managing and Visualizing Spatial Data,
OCEANS 2010, pp1-10.
Li Q et al, A Chunking Method for Euclidean Distance Matrix Calculation on Large Dataset Using
Multi-GPU, IEEE Machine Learning and Applications (ICMLA) Dec 2010, pp 208-213.
Hyo-Sung A, Command Coordination In Multi-Agent Formation: A Distance Matrix Approach,
IEEE Control Automation and Systems (ICCAS) 2010, pp 1592-1597.
Srinkanthan S, Accelerating the Euclidean Distance Matrix Computation Using GPUs, IEEE
Electronics Computer Technology (ICECT) 2011, pp422-426.
Del Bue A et al, 2D-3D Registration of Deformable Shapes with Manifold Projection, 2009, pp
1061-1064.
Drineas P et al, Distance Matrix Reconstruction from Incomplete Distance Information for Sensor
Network Localization, Sensor and Ad Hoc Communications and Networks 2006, pp 536-544.
Ayhan S et al Implementing Geospatially Enabled Aviation Web Services, Integrated
Communications, Navigation and Surveillance (ICNS) 2008, pp1-8.
Xiao Q et al,Application of Visualization Technology in Spatial Data Mining, Computing, Control
and Industrial Engineering (CCIE), 2010, pp153-157.
Rankin J, Visualization of General Defined Space Data, International Journal of Computer
Graphics and Animation, October 2013, Vol3, Nr 4.http://airccse.org/journal/ijcga/current2013.html
Youldash MM & Rankin J, Evaluation of A Geometric Approach ToCoordinatization of Metric
Spaces, 8th International Conference on Information Technology and Applications, ICITA 2013,
July Sydney 2013.

26